Predicting High-LTV Users with ML

The Chart That Silenced a Room

It was a Thursday afternoon in July. Our finance lead, Priya, had been quiet through most of the revenue review. Then she put up one chart — not a deck, just one slide — and the room stopped.

It showed that 2.3% of our 40,000 registered users were generating 71% of all revenue. The other 97.7% — the people we'd spent months and crores acquiring — were effectively break-even or negative.

We weren't running a growth business. We were running an expensive lottery. And we'd been optimizing the lottery ticket price, not the odds.

I remember sitting there thinking: we don't need more users. We need to figure out who those 2% are — before we spend money on them.

What the Data Said

Two Very Different Businesses Hiding in One Dashboard

The problem with averages is they lie gracefully. Our average revenue per user was ₹18/month. Respectable, right? Not when you disaggregate it.

Segment	Users	Avg Monthly Revenue	Revenue Share
All Registered Users	40,000	₹18	—
Monthly Active Users	8,200 (20.5%)	₹87	—
Revenue-Generating Users	920 (2.3%)	₹1,390	71%
Everyone else	39,080 (97.7%)	₹5.40	29%

₹18 average. ₹1,390 reality. Decisions made on blended averages are decisions made on fiction. We'd been optimizing the wrong number for six months.

The marketing team's instinct was to improve onboarding and run re-engagement campaigns. Both felt right. Both were wrong — because they assumed the problem was activation, not selection.

Behavioral Segmentation

What High-LTV Users Did in the First 72 Hours

We pulled 18 months of user event logs and manually tagged every user as High-LTV or not. Then we looked backwards: what did the 2.3% do in their first 3 days that the other 97.7% didn't?

Cohort	72-hr Behavior Pattern	90-Day LTV	D30 Retention
Explorer (never converts)	1 session, bounces	₹0	3%
One-and-Done	1 transaction, disappears	₹45	8%
Occasional User	3+ sessions, 2 features	₹380	24%
High-LTV Power User	5+ sessions, 4+ features, D2-3 return	₹1,390	79%

The Power Users didn't become valuable slowly. They showed every signal of it within 72 hours of signup — they just had no idea they were being watched. The question was whether we were reading the signals in time to act.

One pattern was particularly striking: High-LTV users almost always returned on Day 2 or Day 3. Not Day 7. Not "whenever they remembered." Within 72 hours. That return visit was the single strongest predictor we found.

The Shift

Rethinking the Problem Entirely

The team's first response was "let's improve onboarding so more users behave like the 2%." I pushed back hard on this. You can't engineer someone into becoming a high-intent user through better UX. You can make it easier for them to express that intent — but you can't manufacture it.

The 2% weren't special because of our product. They were special because of who they were when they arrived. Our job was to find more of them — and recognize them faster.

The reframe: stop trying to convert everyone, start predicting who's worth investing in — at the 72-hour mark, not after 30 days of guessing.

The Framework

The Lifetime Value Signal Architecture

We called it LVSA — not because the name matters, but because it helped the team think about it as a system rather than a model. Four layers, all dependent on each other.

Signal Collection

Real-time event capture: session depth, feature sequencing, return timing, comparison actions. Each event assigned a predictive weight based on correlation with 90-day LTV — not based on recency or instinct.

72-Hour Scoring

XGBoost model trained on 18 months of data. Score computed at the 72-hour mark: 0–100. ≥65 = High-LTV candidate. Threshold chosen after testing 60, 65, and 70 — 65 gave the best precision-recall tradeoff for our cost of white-glove treatment.

Experience Routing

Score ≥65: white-glove onboarding, priority support, early premium access. Score ≥80: VP-level manual outreach within 4 hours. Everyone else: standard retention flow. Same product, different investment level.

Monthly Model Refresh

Model retrained monthly with actual 90-day LTV outcomes. This part most teams skip — and it's why their models degrade. Each cycle we improve; each cycle the prediction gets sharper. Score drift alerts trigger manual review.

Behavioral Signal	Predictive Weight	Why
Return visit within 72 hours	0.28	Strongest single predictor of retention
4+ features used in session 1	0.22	Depth of exploration = serious intent
Comparison action performed	0.18	Decision-stage behavior
Session duration > 8 minutes	0.15	Engagement quality
Share or export action	0.10	Product-as-tool signal
Push notification opt-in	0.07	Communication willingness

System Design

How It Was Built

The engineering side took 6 weeks. Not because it was technically complex — it wasn't — but because we had to fight about what it should and shouldn't do for most of Week 1.

Data Pipeline

→Event stream: Segment → Kafka → real-time feature store
→26-feature vector built per user, updated every 2 hours for first 72 hours
→LTV probability score computed at the 72-hour mark and stored in user profile

Product Routing

→High-LTV flag surfaced in internal CRM and support tooling
→White-glove email sequence triggered automatically for ≥65 scorers
→Manual review queue for ≥80 scorers — VP outreach within 4 hours

Acquisition Impact

→Lookalike audiences built from High-LTV cohort behavioral profile (not just demographics)
→CAC ceiling raised 3× for campaigns targeting high-LTV lookalikes
→Source-level LTV attribution: organic had 2.1× more High-LTV users than paid social

Feedback Loop

→Monthly model refresh with actual 90-day revenue outcomes
→Score drift monitoring weekly — any sudden change triggers investigation
→A/B test: white-glove vs standard for borderline scores (60–65)

Execution

90 Days, in Honest Detail

Weeks 1–4 · The Signal Hunt

Finding What Actually Predicted LTV

→ Signal hypothesis validated. Confidence to build the model.

✓Pulled 18 months of user logs; manually labeled 12,000 users as High/Low LTV
✓Built correlation heatmap of 60+ behavioral signals vs actual 90-day revenue
✓Narrowed to 6 signals with combined AUC of 0.78 — not perfect, but actionable

→ Signal hypothesis validated. Confidence to build the model.

Weeks 5–8 · Building and Testing

Model Training and Threshold Calibration

→ Model live in staging. First live scoring run on new signups.

✓XGBoost trained: 18 months, 12K labelled users, 26 features
✓Threshold tested at 60, 65, 70 — 65 gave best precision-recall tradeoff
✓81% precision, 74% recall at threshold 65

→ Model live in staging. First live scoring run on new signups.

Weeks 9–12 · Routing Live

White-Glove Treatment for High Scorers

→ Revenue per acquired user +61% in first cohort

✓High-LTV routing triggered: email sequence, priority support, early premium access
✓Lookalike audiences pushed to Meta and Google from High-LTV behavioral profiles
✓CAC ceiling raised 3× for High-LTV-targeted campaigns

→ Revenue per acquired user +61% in first cohort

Month 4+ · Compounding

Optimization and Source Attribution

→ Blended CAC down 34% without reducing acquisition volume

✓Monthly model refresh established as standard process
✓Source-level LTV attribution added — organic 2.1× better than paid social
✓3 underperforming paid channels paused based on LTV data, not just CAC

→ Blended CAC down 34% without reducing acquisition volume

Results

What Actually Changed

+61%

Revenue per Acquired User

first cohort after routing

−34%

Blended CAC

without volume reduction

81%

Model Precision

at 65-score threshold

2.7×

High-LTV User Acquisition

YoY vs pre-model baseline

Metric	Before	After (90 Days)
High-LTV User Share of New Cohort	2.3%	6.1%
Revenue per Acquired User	₹18/mo avg	₹29/mo avg
Blended CAC	₹340	₹225
Paid Channel ROAS	1.8×	3.4×
Model Precision	—	81%

The number I'm proudest of isn't the revenue lift. It's that the High-LTV share of new cohorts went from 2.3% to 6.1%. We didn't get lucky more often — we got better at recognizing luck when it arrived.

Deep Insights

What I'd Tell Someone Starting This Today

LTV is a sorting problem disguised as a growth problem

Most teams try to grow LTV by adding features, improving onboarding, or running engagement campaigns. The real lever is routing: get high-LTV users to intensive engagement early; let low-LTV users self-serve. The product already existed. The routing didn't.

The 72-hour window matters more than any other

Behavior in the first 3 days after signup predicted 90-day LTV better than any demographic, channel, or survey response. High-value users show themselves almost immediately — if you're looking.

Set different CAC ceilings by predicted LTV segment

If you apply one CAC ceiling across all users, you systematically underspend acquiring high-LTV users and overspend on low-LTV ones. Differentiated CAC limits restructure the economics of every channel you run.

Precision beats recall when white-glove treatment has real costs

At 81% precision and 74% recall, we deliberately accepted missing some true High-LTV users. The cost of treating false positives (low-LTV users) to white-glove standards was higher than missing true positives. Know your cost structure before choosing your threshold.

The model output becomes your best acquisition brief

Once you know exactly who your high-LTV users are at a behavioral level, building lookalike audiences is almost automatic. The model didn't just improve retention investment — it improved acquisition targeting simultaneously.

What Didn't Work

The Mistakes I'm Glad We Made Early

✕ Our first model had 83 features and overfit badly

Why it broke

Validation AUC looked great at 0.84. Live scoring performance degraded within 3 weeks — the model had memorized training noise. We were celebrating the wrong metric.

What fixed it

Cut to the top 10 features by mutual information score. Validation AUC dropped to 0.78, but live scoring AUC improved to 0.81. Less is genuinely more when the signal-to-noise ratio is this low.

✕ We started white-glove treatment at score ≥60

Why it broke

Felt conservative enough. It wasn't. 40% of users above 60 churned within 30 days — too many false positives getting expensive treatment.

What fixed it

Raised threshold to 65. False positive rate dropped 51%. Manual VP outreach reserved for ≥80. At that threshold, paid-to-conversion was 3.4×.

✕ We forgot to surface the score to the sales team

Why it broke

High-LTV users were being routed to the right email sequences but handled identically to everyone else in sales calls. Reps had no visibility into their score.

What fixed it

LTV score added to CRM view. Tiered playbooks by score band. Score ≥80 users got VP-level call within 4 hours of signup. That change alone moved conversion rate for those users by 28%.

Where This Goes

The Next Version of This

Real-Time Score Updates

Moving from 72-hour static scoring to continuous — recalculated every 6 hours. Some users cross the threshold on Day 4 or 5. We're currently missing them.

Score Velocity Triggers

It's not just the score — it's the direction and speed of change. A user going from 45 to 72 in 48 hours is different from one who's been at 73 for 5 days. Velocity should trigger different interventions.

Churn-Weighted LTV Scoring

Incorporate churn probability into the score. High LTV with high churn risk needs a different intervention than stable mid-LTV. Combining both would create a genuinely actionable risk-adjusted view.

We didn't need more users. We needed to get much better at recognizing the right ones the moment they walked in the door.

How We Used ML to Predict Insurance Buyers (50,000+ Data Points)

Machine Learning

Building a Real-Time Intent Scoring Engine for Lead Conversion

Machine Learning

Building a Real-Time Intent Scoring Engine for Lead Conversion

9 min read

Machine Learning

Using ML to Reduce CAC by 31% via Smart Audience Segmentation

8 min read

Machine Learning

Personalizing Insurance Offers Using ML-Driven User Clustering

8 min read

All Case Studies

Browse all 30 breakdowns

Work With Me

Let's build something together

Connect on LinkedIn

Follow for more insights

Where to go next

Product Monetisation Consultant
The consulting work this sits inside, with the rest of the evidence behind it.
Average revenue per user benchmark, subscription apps
Published figures, traced to their sources — and where the sources disagree.
Trial-to-paid conversion benchmark, subscription apps
Published figures, traced to their sources — and where the sources disagree.
Product Growth Score
Thirteen questions that locate which part of the funnel is actually constraining growth.

How We Predicted High-LTV Users Using Behavioral ML (100K+ Sessions)

Results at a glance

The Chart That Silenced a Room

Two Very Different Businesses Hiding in One Dashboard

What High-LTV Users Did in the First 72 Hours

Rethinking the Problem Entirely

The Lifetime Value Signal Architecture

How It Was Built

90 Days, in Honest Detail

What Actually Changed

What I'd Tell Someone Starting This Today

LTV is a sorting problem disguised as a growth problem

The 72-hour window matters more than any other

Set different CAC ceilings by predicted LTV segment

Precision beats recall when white-glove treatment has real costs

The model output becomes your best acquisition brief

The Mistakes I'm Glad We Made Early

The Next Version of This

You might also like

Where to go next

The Chart That Silenced a Room

Two Very Different Businesses Hiding in One Dashboard

What High-LTV Users Did in the First 72 Hours

Rethinking the Problem Entirely

The Lifetime Value Signal Architecture

How It Was Built

90 Days, in Honest Detail

What Actually Changed

What I'd Tell Someone Starting This Today

LTV is a sorting problem disguised as a growth problem

The 72-hour window matters more than any other

Set different CAC ceilings by predicted LTV segment

Precision beats recall when white-glove treatment has real costs

The model output becomes your best acquisition brief

The Mistakes I'm Glad We Made Early

The Next Version of This